Research · Accepted conference paper
5G Latency Violation Prediction
Forecasting 5G eMBB latency breaches five minutes ahead.
In plain words. Research on warning a mobile network a few minutes before it slows down, so operators can react earlier. It was tested on a controlled lab network.
The problem
Threshold alarms fire after a delay has already degraded the user experience.
What it does
Compares sequence models (LSTM, dilated TCN) with an engineered-feature tabular baseline and a stacked ensemble. The target is whether average latency over the next ten 30-second intervals exceeds a threshold. A real-time predictor ingests live ping and iperf3 measurements from the lab network.
Key decisions
- Features are shifted before rolling aggregation and preprocessing is fit only on training data, to prevent leakage.
- Chronological evaluation instead of shuffled windows.
- Reported with PR-AUC and MCC because violations are the minority class.
Limits and status
- Tested on a controlled lab network, not on a live carrier deployment.
- Feature attribution is not causal proof.